查看更多>>摘要：The quest to find reliable biomarkers in autism spectrum disorders (ASD) is an ongoing endeavour to identify both underlying causes and measurable indicators of this neurodevelopmental condition. Machine learning (ML) and advanced deep learning (DL) techniques have enhanced biomarker identification in neuroimaging and behavioral studies, aiding in diagnostic accuracy and early detection. This review paper examines the transformative impact of applying machine learning (ML), particularly deep learning (DL) techniques such as transfer learning and transformer architectures, in advancing ASD diagnosis. The review begins by critically assessing existing literature utilizing ML techniques like logistic regression, random forest, and support vector machines in identifying biomarkers that could potentially aid in the diagnosis of ASD and differentiate between ASD and neurotypical individuals. The focus then shifts to DL models, including Multilayer Perceptrons, Convolutional Neural Networks, Graph Neural Networks, and Long Short-Term Memory networks, to evaluate their suitability for identifying complex patterns linked to ASD. Addressing limited datasets, the review examines transfer learning with pre-trained models, including VGG, ResNet, DenseNet, MobileNet, Inception, and Xception architectures. Additionally, using the ABIDE-I dataset, VGG19, MobileNet, InceptionV3, and DenseNet121 were applied, evaluating their performance through accuracy, sensitivity, specificity, and F1 score. The review further considers transformer architectures, such as Vision Transformers, Swin Transformers, Spatial Temporal Transformers, BolT Transformer, and Convolutional Network Transformer for capturing longrange dependencies in ASD diagnosis. This review aims to be an essential reference for researchers exploring the field of AI-powered ASD diagnosis and classification, by offering analysis of various approaches and highlighting recent advancements.
NETL
NSTL
Elsevier